Cocktails

Unknown — Thu, 13 Nov 2025 00:00:00 +0000

Making cocktails at home is good because:</p>

You can customize them to your or your guests’ liking</li>
You can make cocktails with high-end liquors that you pick out</li>
You can enjoy drinking in your own quiet space</li>
Bartending at home is cost effective</li> </ol>
I am by no means an expert on this topic but here is what I’ve learned so far.</p>
Specs</a></h2>
These aren’t the “true” “official” recipes, but rather my own hacked together versions that I like.</p>
Old Fashioned</a></h3>
There are many variations and you can use any spirit, although whiskey is the classic.</p>
</figure>

Ingredients</dt>

2oz whiskey</li>
1/8oz–1/4oz simple syrup</li>
2 dashes Angostura bitters</li>
2 dashes orange bitters</li> </ul> </dd>
Method</dt>
Stir in mixing glass with cubed ice. Strain. Pour over ice.</dd>
Serving and Garnish</dt>
Rocks glass, large ice cube, Luxardo cherry, orange peel twist</dd>
Good Bottles</dt>
Four Roses, Maker’s 46. Diplomatico Reserva</dd>
Riffs</dt>

Swap spirits</em>: 2oz aged rum for whiskey</li>
Swap sweetener</em>: swap demerara / maple / blue agave syrup for simple syrup</li>
Add coffee depth</em>: 1/2oz coffee liqueur</li>
Exotic bitters</em>: choose — 2 dashes Aztec chocolate bitters, 2 dashes black walnut bitters</li>
Port old fashioned</em>: 2oz bourbon, 3/4oz 10-year Tawny port, 1/4oz maple syrup, 2 dashes Angostura bitters</li> </ul> </dd> </dl>
Whiskey Sour</a></h3>
</figure>

Ingredients</dt>

2oz whiskey / rum</li>
1/2oz maple / 2:1 demerara syrup OR</strong> 3/4oz 1:1 simple syrup</li>
3/4oz lemon juice</li>
(Optional) Fee Foam or 1 egg white</li> </ul> </dd>
Method</dt>
Shake hard with large ice cube (for max foam). Double strain into glass.</dd>
Serving and Garnish</dt>
Rocks glass (over one large ice cube) or cocktail glass (up). Garnish with a Luxardo cherry and a few drops of Angostura bitters on top of the foam.</dd>
Good Bottles</dt>
Four Roses, Wild Turkey 101</dd>
Riffs</dt>

Quick version with lemonade</em>: Swap lemon juice and sweetener for 2–3oz of fresh unpasteurized lemonade + a quarter of a lemon, juiced fresh. The Trader Joe’s refrigerated lemonade works perfectly, no need for additional sweetener.</li>
New York Sour</em>: Float 1/2oz dry red wine (e.g. Merlot, Malbec) or port on top over the back of a barspoon</li> </ul> </dd> </dl>
Manhattan</a></h3>
</figure>

Ingredients</dt>

2oz whiskey</li>
1oz sweet vermouth (or ruby port)</li>
2 dashes Angostura bitters</li>
(Optional) 1 dash black walnut bitters</li> </ul> </dd>
Method</dt>
Stir in mixing glass with cubed ice. Strain.</dd>
Serving and Garnish</dt>
Coupe glass, up, Luxardo cherry</dd>
Good Bottles</dt>
Carpano Antica, Carpano Punt E Mes, Cocchi Vermouth Di Torino. Otima 10 Year Port. Rittenhouse 100.</dd>
Riffs</dt>

Perfect Manhattan</em>: Use split of 1/2oz sweet vermouth + 1/2oz dry vermouth</li>
Perfect Summit Manhattan</em></a>: 2 1/2oz bourbon, 3/4oz dry vermouth, 3/4oz Muscat wine, 1/6oz curacao, 2 dashes bitters</li>
Black Manhattan</em>: 2oz bourbon, 1oz Amaro Averna (or other Amaro such as Montenegro, Ramazzotti, or Nonino), 2 dashes Angostura bitters, 2 dashes orange bitters</li>
Monte Carlo</em>: 2oz rye, 1/2oz Benedictine, 1-2 dashes Angostura bitters</li>
Winter Holiday</em></a>: 2oz rye, 3/4oz tawny port, 1/4oz dark creme de cacao, 1 dash Angostura bitters, 1 dash chocolate bitters</li> </ul> </dd> </dl> Whiskey Highball</a></h3>
</figure>

Ingredients</dt>

2oz whiskey / rum / gin</li>
4–6oz mixer (club soda, ginger ale, coke, sprite)</li>
1/4oz–1/2oz fresh lime/lemon juice</li> </ul> </dd>
Method</dt>
Add ice to glass and stir until glass is frosted. Pour out melted water. Pour in spirit and citrus and top with the mixer. Stir once. Note</em>: some suggest to pour mixer first, then spirit — experiment with it.</dd>
Serving and Garnish</dt>
Highball glass, on rocks. Luxardo cherry, lemon twist.</dd>
Good Bottles</dt>
Suntory Toki.</dd>
Note on Mixers</dt>

Sparkling water</em>: filtered, sometimes un-mineralized, tap water with added carbonation</li>
Spring water with added carbonation</em>: mineralized still spring water + added carbonation (e.g. San Pellegrino, Mountain Valley)</li>
Naturally sparkling spring water</em>: naturally carbonated and mineralized water from a spring, often with some added carbonation (e.g. Topo Chico, Perrier, Gerolsteiner)</li>
Club soda</em>: filtered tap water + carbonation + sodium bicarbonate (adds a slight salty flavor) (e.g. Fever-Tree Club Soda)</li>
Tonic water</em>: filtered tap water + carbonation + quinine (slightly bitter taste) + sugar/sweeteners (can contain HFCS, offsets bitterness from quinine)</li>
Seltzer</em>: be careful! This actually means nothing today. Identify which of the above categories the water actually falls in.</li>
For highballs</em>: club soda is often the best choice since the slight bit of saltiness enhances the spirit’s flavor</li> </ul> </dd>
Riffs</dt>

Tom Collins</em>: 2oz gin, 1oz lemon juice, 1/2–1oz simple syrup, club soda to top in highball glass. Garnish with lemon wheel and a maraschino cherry.</li>
Gin and Tonic</em>: 2oz gin, 2oz tonic water, a squeeze of lime, and a dash of grenadine. Hendrick’s gin is nice.</li> </ul> </dd> </dl>
Daiquiri</a></h3>
</figure>

Ingredients</dt>

2oz white rum</li>
3/4oz lime juice</li>
3/4oz simple syrup</li> </ul> </dd>
Method</dt>
Shake very hard with ice for a full 15 seconds. Double strain into glass.</dd>
Serving and Garnish</dt>
Coupe or Nick and Nora glass (chill the glass in the freezer), up</dd>
Good Bottles</dt>
El Dorado 3, Probitas, Planteray 3 Stars, Flor de Cana 4</dd>
Riffs</dt>

Herbal enhancement</em>: add a dash of absinthe before shaking</li>
Grassy funk enhancement</em>: according to Cocktail Codex, the ideal daiquiri is 1 3/4oz white rum + 1/4oz rhum agricole</em> + a full 1oz of lime juice</li>
Difford’s version: Daiquiri No. 1</em>: 1 2/3oz filtered white rum, 1/2oz lime juice, 1/3oz 2:1</em> sugar syrup (thicker syrup prevents overdilution)</li> </ul> </dd> </dl>
El Presidente</a></h3>
</figure>

Ingredients</dt>

1 1/2oz light rum</li>
3/4oz blanc or dry vermouth</li>
1/4oz orange liqueur</li>
1 barspoon real grenadine</li> </ul> </dd>
Method</dt>
Stir in mixing glass with ice. Strain.</dd>
Serving and Garnish</dt>
Cocktail glass, up</dd>
Good Bottles</dt>
Diplomatico Planas, Probitas, El Dorado 3, Planteray 3 Stars. Lillet Blanc, Dolin Vermouth Blanc.</dd> </dl>
Mai Tai</a></h3>
</figure>

Ingredients</dt>

2oz rum</li>
1oz lime juice</li>
1/2oz dry curacao</li>
1/2oz orgeat</li>
1/8–1/4oz simple / demerara syrup</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Large rocks glass, ice pebbles, lime peel</dd>
Good Bottles</dt>
Appleton Estate 8 / 12 / 15, El Dorado 8, Smith and Cross. Pierre Ferrand Dry Curacao.</dd>
Riffs</dt>

“Hotel” Mai Tai</em></a> (from the NoMad Cocktail Book): small lime wedge, 1oz lime juice, 3/4oz orgeat, 1/2oz creole shrubb (orange liqueur), 1/2oz rhum agricole (Rhum J.M Blanc Agricole Martinique 100 Proof), 3/4oz Appleton Signature (4 year blend), 3/4oz El Dorado 15 (highly aged dark rum)</li> </ul> </dd> </dl> Gimlet</a></h3>
</figure>

Ingredients</dt>

2oz gin</li>
1/2–3/4oz lime juice</li>
1/2–3/4oz simple syrup</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Nick and nora or coupe, up, lime wheel garnish</dd>
Good Bottles</dt>
Citadelle, Hendrick’s, Tanqueray No. Ten, Tanqueray Rangpur, Bluecoat, Monkey 47, Junipero, Barr Hill, Isle of Harris, Roku Gin (Japanese), Gin Mare (Spanish, exotic, olives), Empress 1908 Indigo</dd>
Riffs</dt>

Gimlet highball</em>: 1 1/2oz gin, 1/2oz lime juice, 1/4–1/2oz simple syrup. Shake and pour. Top with 3–4oz club soda.</li> </ul> </dd> </dl>
Negroni</a></h3>
</figure>

Ingredients</dt>

1oz gin</li>
1oz sweet vermouth</li>
1oz Campari</li> </ul> </dd>
Method</dt>
Stir in mixing tin with ice. Strain.</dd>
Serving and Garnish</dt>
Rocks glass, one large ice cube, orange peel garnish</dd>
Good Bottles</dt>
Cocchi Vermouth di Torino, Cinzano Rosso Vermouth, Carpano Antica Formula, Dolin Rouge Sweet. Tanqueray No. 10.</dd>
Riffs</dt>

MonteNegroni</em>: swap Campari for Amaro Montenegro. 2oz Amaro Montenegro, 1oz gin, 1oz sweet vermouth, 2 dashes Angostura bitters.</li>
Kingston Negroni</em>: 1oz Jamaican rum (Smith and Cross), 1oz sweet vermouth, 1oz Campari</li> </ul> </dd> </dl>
Martini</a></h3>
</figure>

Ingredients</dt>

2oz gin</li>
1/2–3/4oz dry vermouth</li>
1 dash orange bitters</li> </ul> </dd>
Method</dt>
Stir in mixing tin with ice. Strain. (Optionally shake)</dd>
Serving and Garnish</dt>
Martini glass, up, lemon peel / green olive garnish</dd>
Good Bottles</dt>
Plymouth, Bluecoat. Dolin Dry, Noilly Prat Extra Dry.</dd>
Riffs</dt>

Perfect Martini</em>: add 1/2–3/4oz sweet vermouth to the tin</li> </ul> </dd> </dl>
Lemon Drop</a></h3>
</figure>

Ingredients</dt>

1 1/2oz gin</li>
3/4oz lemon juice</li>
1/2oz triple sec (sub 1oz limoncello + 1/4oz triple sec)</li>
1/4–1/2oz simple syrup</li>
(Optional) one dash orange bitters</li>
(Optional) pinch of salt</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Coupe glass, up. Lemon twist or lemon wheel for garnish.</dd>
Good Bottles</dt>
Plymouth, Tanqueray No. 10.</dd> </dl>
Boulevardier</a></h3>
</figure>

Ingredients</dt>

1 1/2oz bourbon</li>
3/4–1oz Campari</li>
3/4–1oz sweet vermouth</li>
(Optional) small pinch of salt</li> </ul> </dd>
Method</dt>
Stir in mixing tin with ice. Strain. The classic recipe uses 1oz Campari and vermouth, but using 3/4oz Campari and vermouth may be better.</dd>
Serving and Garnish</dt>
Rocks glass or coupe, one large ice cube, orange peel garnish</dd>
Good Bottles</dt>
Same Vermouths as the Negroni. Buffalo Trace, Elijah Craig Small Batch.</dd> </dl>
Sidecar</a></h3>
</figure>

Ingredients</dt>

1 1/2–2oz Cognac / brandy</li>
1/2oz Cointreau</li>
3/4oz lemon juice</li>
1/4–1/3oz simple syrup</li>
Tiny pinch of salt</li> </ul> </dd>
Method:</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Coupe, up. Garnish with an orange or lemon twist.</dd>
Good Bottles</dt>
Pierre Ferrand Ambre, Pierre Ferrand 1840 Original Formula, Hennessy VS, Remy Martin VSOP</dd>
Riffs</dt>

Gennaro’s sidecar</em>: 1oz Cognac, 3/4oz Cointreau, 1/2oz limoncello, 1/2oz lemon juice</li>
Champs-Elysees</em>: 2oz Cognac, 1/2oz Green Chartreuse, 3/4oz lemon juice, 1/4–1/2oz simple syrup, 2 dashes Angostura bitters</li> </ul> </dd> </dl>
Between the Sheets</a></h3>
</figure>

Ingredients</dt>

1oz cognac</li>
1oz light rum</li>
1oz triple sec (Cointreau)</li>
1/4oz lemon juice</li>
(Optional) 1/4oz simple syrup</li>
(Optional) pinch of salt</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Cocktail glass, up, express orange peel (no garnish)</dd>
Good Bottles</dt>
Same cognacs as above. Planteray 3 Stars, Flor De Cana 4, El Dorado 3, Diplomatico Planas.</dd> </dl>
Mudslide</a></h3>
</figure>

Ingredients</dt>

1oz Irish cream</li>
1oz coffee liqueur</li>
1oz vanilla vodka</li>
1 1/2oz heavy cream</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Rocks glass or coupe, fresh ice OR up, chocolate syrup. Garnish with shaved chocolate.</dd>
Good Bottles</dt>
Baileys, Kahlua</dd> </dl>
Port of Call</a></h3>
</figure>

Ingredients</dt>

1oz gin</li>
1oz ruby port wine</li>
3/4oz lemon juice</li>
1/2oz cinnamon syrup</li>
1 tsp cranberry preserves</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Rocks glass, crushed ice, mint garnish</dd>
Good Bottles</dt>
I have no idea</dd>
Riffs</dt>

If cinnamon syrup / cramberries aren’t available, sub with: 1/2oz simple syrup + 1 dash Angostura bitters + dash of grenadine</li> </ul> </dd> </dl>
Japanese Snow Monkey</a></h3>
I had this cocktail</a> at Rumba</a>. The spec is given as: “Kiyomi rum, Nigori sake, Bermutto dry vermouth, Campari, raspberry, and lemon”. “A floral but punchy take on a Negroni to help brighten long winter days.” There was a particular roasted corn-like and slightly smoky flavor that was hard to pin down, but was very good. My attempted reproduction:</p>
</figure>

Ingredients</dt>

1oz Japanese rum (substitute with: 1/2oz Probitas + 1/4oz El Dorado 3 + 1/4oz agricole blanc)</li>
1oz Nigori sake</li>
3/4oz dry vermouth</li>
1/2oz Campari</li>
1/2oz raspberry syrup OR</strong> muddle 6-8 raspberries with 1/4oz simple syrup</li>
1/2oz lemon juice</li>
Sprinkle of salt</li>
(Optional) 3-4 drops of mezcal (for slight smoke)</li> </ul> </dd>
Method</dt>
Muddle raspberries. Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Coupe, up.</dd>
Good Bottles</dt>
Probitas. Tozai Snow Maiden, Sho Chiku Bai Nigori.</dd> </dl>
Last Word</a></h3>
</figure>

Ingredients</dt>

3/4oz gin</li>
3/4oz Green Chartreuse</li>
3/4oz Maraschino liqueur</li>
3/4oz lime juice</li> </ul> </dd>
Method</dt>
Shake with ice for 20-30 seconds. Double strain.</dd>
Serving and Garnish</dt>
Coupe / nick and nora, up. Garnish with Luxardo cherry + lime twist.</dd>
Riffs</dt>

It seems widely recognized that it is better to go lower on Maraschino and up on gin for better balance: 1oz gin + 1/2oz Maraschino liqueur.</li> </ul> </dd> </dl>
Bee’s Knees</a></h3>
</figure>

Ingredients</dt>

2oz gin</li>
3/4oz honey syrup (equal parts honey and warm water, stir until mixed and cool)</li>
3/4oz lemon juice</li>
(Optional) splash of fresh lemonade</li> </ul> </dd>
Method</dt>
Shake with ice for 15 seconds. Double strain.</dd>
Serving and Garnish</dt>
Coupe / nick and nora, up. Garnish with lemon twist.</dd> </dl>
Jasmine</a></h3>
</figure>

Ingredients</dt>

1 1/2oz gin</li>
1/4oz Campari</li>
1/4oz orange liqueur (Cointreau)</li>
3/4oz lemon juice</li>
1/8–1/4oz simple syrup</li> </ul> </dd>
Method</dt>
Shake with ice for 15-20 seconds. Double strain.</dd>
Serving and Garnish</dt>
Coupe, up. Garnish with lemon twist.</dd> </dl>
Monkey Gland</a></h3>
</figure>

Ingredients</dt>

1 1/2oz gin</li>
1oz fresh orange juice</li>
1/4oz grenadine</li>
1 dash absinthe</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Martini glass or coupe, up. Garnish with orange twist or wheel.</dd> </dl>
Corpse Reviver #2</a></h3>
</figure>

Ingredients</dt>

3/4oz gin</li>
3/4oz orange liqueur (Cointreau)</li>
3/4oz Lillet Blanc</li>
3/4oz lemon juice</li>
1 dash absinthe for rinsing the glass</li> </ul> </dd>
Method</dt>
Spray the glass with absinthe. Shake ingredients with ice for 15 seconds. Double strain.</dd>
Serving and Garnish</dt>
Coupe, up. Garnish with lemon or orange twist.</dd> </dl>
Martinez</a></h3>
</figure>

Ingredients</dt>

1 1/2oz gin</li>
1 1/2oz sweet vermouth</li>
1/4oz Luxardo maraschino liqueur</li>
2 dashes Angostura bitters</li> </ul> </dd>
Method</dt>
Stir over ice until chilled, 30-45 seconds. Strain.</dd>
Serving and Garnish</dt>
Coupe, up. Garnish with lemon twist.</dd> </dl>
Bijou</a></h3>
</figure>

Ingredients</dt>

1 1/2oz gin</li>
1oz sweet vermouth</li>
3/4oz Green Chartreuse</li>
2 dashes orange bitters</li> </ul> </dd>
Method</dt>
Stir over ice until chilled, 20-30 seconds. Strain.</dd>
Serving and Garnish</dt>
Nick and nora, up. Garnish with cherry.</dd> </dl>
Aviation</a></h3>
</figure>

Ingredients</dt>

1 3/4oz gin</li>
1/2oz Luxardo Maraschino liqueur</li>
1/4oz Creme de Violette Liqueur</li>
1/2oz lemon juice</li>
1/4oz cold water (omit if using wet ice)</li> </ul> </dd>
Method</dt>
Shake with ice until chilled, 10-15 seconds. Double strain.</dd>
Serving and Garnish</dt>
Nick and nora or martini glass, up. Garnish with cherry or lemon twist.</dd> </dl>
Paper Plane</a></h3>
</figure>

Ingredients</dt>

3/4oz bourbon</li>
3/4oz Aperol</li>
3/4oz Amaro Montenegro (typically Amaro Nonino is used)</li>
3/4oz lemon juice</li> </ul> </dd>
Method</dt>
Shake with ice until chilled, 10-15 seconds. Double strain.</dd>
Serving and Garnish</dt>
Nick and nora or coupe glass, up.</dd>
Riffs</dt>

Use a different amaro (e.g. Nonino, Averna)</li> </ul> </dd> </dl>
Spritz</a></h3>
</figure>

Ingredients</dt>

3oz dry sparkling wine (e.g. Prosecco)</li>
2oz amaro (e.g. Aperol, Amaro Montenegro, Campari)</li>
1oz soda water</li>
(Optional) 1/4oz lemon juice</li> </ul> </dd>
Method</dt>
Use the classic 3-2-1 spritz formula and improvise the ingredients! In the glass, first add wine, then slowly drop in ice, then amaro, and finally soda water. Stir once.</dd>
Serving and Garnish</dt>
Serve in a wine glass. Garnish with a slice of orange.</dd>
Riffs</dt>

Hugo spritz</em>: Muddle 8-10 mint leaves with 1 lime wedge and 1.5oz Elderflower liqueur. Top with ice. Pour 2–3oz Prosecco + 2oz sparkling water.</li> </ul> </dd> </dl>
Gin Fix</a></h3>
</figure>

Ingredients</dt>

2oz gin</li>
1/2oz simple syrup</li>
1oz lemon juice</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Highball or rocks glass, fill glass with crushed ice. Stir after pouring. Garnish with lemon peel.</dd> </dl>
White Negroni</a></h3>
</figure>

Ingredients</dt>

1 1/2oz gin</li>
1oz Lillet Blanc</li>
1/2oz Suze</li> </ul> </dd>
Method</dt>
Stir over ice. Strain.</dd>
Serving and Garnish</dt>
Rocks glass. Garnish with a lemon twist.</dd>
Riffs</dt>

Alpine Negroni</em>: 1oz gin, 3/4oz Cocchi Americano, 3/4oz Suze, 1/2oz Dolin Genepy, 1 barspoon White Creme de Menthe, 3 dashes lemon bitters, sprinkle of salt.</li> </ul> </dd> </dl>
Elk’s Own</a></h3>
</figure>

Ingredients</dt>

1oz whiskey</li>
3/4oz Ruby Port</li>
3/4oz lemon juice</li>
1/2oz demerara syrup</li>
1–2 dashes chocolate / orange bitters</li>
Egg white / Fee foam</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Nick and Nora glass.</dd> </dl>
Jungle Bird</a></h3>
</figure>

Ingredients</dt>

1 1/2oz dark rum</li>
1/2oz Campari (or Aperol for less bitterness)</li>
1 1/2oz pineapple juice</li>
1/2oz lime juice</li>
1/2oz simple syrup</li> </ul> </dd>
Method</dt>
Shake hard with pebble ice. Ungated pour into glass.</dd>
Serving and Garnish</dt>
Fill a rocks glass with pebble ice and pour. Garnish with slices of lime and orange.</dd>
Good Bottles</dt>
Rhum Barbancourt, Smith and Cross, Appleton Estate 8/12</dd> </dl>
French 75</a></h3>
</figure>

Ingredients</dt>

1oz gin</li>
1/2oz simple syrup</li>
1/2oz lemon juice</li>
3–4oz Brut Champagne / dry sparkling wine</li> </ul> </dd>
Method</dt>
Shake gin + syrup + lemon juice with ice. Pour into flute. Top off with Champagne.</dd>
Serving and Garnish</dt>
Serve in Champagne flute, up. Garnish with lemon twist</dd>
Good bottles</dt>
Empress 1908 Elderflower Rose Gin, Empress 1908 Indigo Gin, Hendrick’s</dd>
Riffs</dt>

French 125</em>: Sub gin with cognac.</li>
French 95</em>: Sub gin with bourbon.</li>
Sub simple syrup for Elderflower liqueur</li>
Other interesting riffs can be found here</a></li> </ul> </dd> </dl>
Poet’s Dream</a></h3>
</figure>

Ingredients</dt>

3/4oz gin</li>
1/2oz Benedictine</li>
3/4oz dry vermouth</li>
(Optional) touch of dry curacao</li> </ul> </dd>
Method</dt>
Stir on ice. Double strain into chilled glass.</dd>
Serving and Garnish</dt>
Nick and Nora glass, up. Garnish with lemon peel</dd> </dl>
Adriatique</a></h3>
</figure>

Ingredients</dt>

1oz Amaro Montenegro</li>
1/2oz Aperol</li>
1oz orange juice</li> </ul> </dd>
Method</dt>
Shake with ice. Double strain.</dd>
Serving and Garnish</dt>
Nick and Nora glass, up. Orange peel garnish.</dd>
Riffs</dt>

A more balanced version</em>: (1) Add 1/4oz lemon juice to boost the acidity of the orange juice. (2) Replace the Aperol with Campari. (3) Add 1/8oz of simple syrup.</li>
Spritz version</em>: Top off the base cocktail with 3–4oz of dry sparkling white wine (e.g. Prosecco) or seltzer</li> </ul> </dd> </dl>
Americano</a></h3>
</figure>

Ingredients</dt>

1 1/2oz Amaro (e.g. Campari, Amaro Montenegro, Aperol — if sweetness is desired)</li>
1 1/2oz sweet vermouth</li>
2–3oz seltzer</li>
(Optional) 1–2 dashes orange bitters</li> </ul> </dd>
Method</dt>
Pour Amaro and vermouth over ice in glass. Stir until chilled. Top with seltzer. One final stir.</dd>
Serving and Garnish</dt>
Highball glass, filled with large cubed ice. Garnish with orange wheel / lemon twist.</dd> </dl>
Poison Reviews</a></h2>
Here are things I’ve tried and what I think. To be clear</em>: I have no actual taste in alcohol. Total amateur opinions to follow.</p>
Whiskey</a></h3>

Maker’s Mark 46</em>. (6/10). I like it. Good enough neat.</li>
Woodford Reserve (Regular)</em>. (4/10). Pretty good, but I prefer Maker’s 46 over this. Not so great anymore after having better whiskeys.</li>
Suntory Toki</em>. (6/10). I really like it when mixed. Ideal for highballs and sours. Not great neat.</li>
Four Roses Small Batch</em>. (8/10). The best so far. Excellent. Very smooth and flavorful. Very enjoyable neat. Spending more on bourbon (at least $40 for a 750ml bottle) seems to be worth it.</li>
Redbreast 12</em>. TBD.</li>
Michter’s US1</em>. TBD.</li> </ul> Rum</a></h3>

Diplomatico Reserva Exclusiva</em>. (7/10). I like it. A few feel it is too harsh and doesn’t go down smooth, at least neat. But I’m a fan.</li>
El Dorado 3</em>. (6/10). Slight hint of fruity/tropical flavor but overpowered by a pure alcohol aftertaste. Reasonably OK in daiquiris, but not my favorite.</li>
Clement Rhum Agricole Blanc.</em> (9/10). Absolutely essential for a daiquiri. A great flavor enhancer with notes of banana and tropical fruits + sweetness.</li>
Probitas</em>. (9/10). This goes very well in a daiquiri. Strong flavors of ripe banana and pineapple. Strong at 48%: I recommend some explicit dilution of 1/4oz chilled water in a daiquiri + a touch more than 3/4oz simple syrup.</li>
Planteray 3 Stars</em>. TBD.</li>
Appleton Estate 8</em>. TBD.</li> </ul> Soju</a></h3>

Hwayo 25 Soju</em>. (5/10). Has an initial taste much closer to sake than soju, but still inferior to sake. Good to try, but the aftertaste is still too alcoholic.</li>
Ilpoom Jinro Oak 43 Soju</em>. TBD.</li> </ul> Gin</a></h3>

Citadelle Gin</em>. (9/10). Excellent ‘celery’-like + juniper berry flavor, freshness as the final note. Gimlet with this is too good.</li>
Hendrick’s Gin.</em> (8/10). Also very good, but more subtle finish and harsher down the throat vs Citadelle.</li>
Barr Hill Gin</em>. (8/10). This is a very classic gin with mainly juniper notes. There is a hint of honey, but it needs a honeyed cocktail to make it stand out.</li> </ul> Wines</a></h3>
Port</a></h4>

Porto Morgado Old Tawny (10 Years)</em>. (7/10). Sits around $10 from Trader Joe’s. Hits way way above its price point. We love this.</li>
Porto Morgado Ruby</em>. (5/10). Not as good. Similar price point, but much worse than above imo.</li>
Fairbanks Cellars California Port (1.5L)</em>. (2/10). Cheap but bad. Stay away.</li>
Warre’s Otima 10 Tawny Porto</em>. TBD.</li> </ul> Reds</a></h4>

Decoy Cabernet</em>. (5/10). Decent, hits a bit above its price point, but it’s better to spend more.</li>
Sean Thackrey Pleiades XXXII</em>. (9/10) Very good. Fruity and smooth.</li>
Charles Shaw wines (all varieties)</em>. (3/10). Very cheap, but do not buy lol. Get something nicer.</li>
No Es Pituko Carignan 2024</em>. (8/10). So so smooth, hard to believe there was any alcohol at all. Excellent.</li>
Domaine de la Patience, Red (2023)</em>. (8/10). Another smooth and gulpable red.</li>
Love Red 2022 - Broc Cellars</em>. (7/10). Similar to the other smooth red wines. Light flavor, very smooth, hardly any taste of alcohol. I think I need something more flavorful next time.</li>
Chateau Pey-Bonhomme Le-Tours Merlot</em>. TBD.</li>
Folk Machine Pinot Noir</em>. TBD.</li> </ul> Whites</a></h4>

2021 Virgo Magic Wines ‘Disco Lemonade’ Riesling</em>. (5/10). Tastes heavily like brined olives. I think this is an acquired taste, not for a general audience.</li>
Barefoot Moscato</em>. (2/10). Cheap, mild tasting, and gulpable. Get something nicer.</li>
Risata Moscato d’Asti DOCG</em>. (8/10). Low alcohol, smooth, sweet, and fruity. A crowd favorite.</li>
2023 Nickel & Nickel Grand Sky Vineyard Chardonnay</em>. (9/10) Incredibly good, smooth, bright lime aftertaste.</li>
2024 Elio Perrone Moscato d’Asti Sourgal</em>. (6/10). Sparkling, sweet, honeyed, light. A bit too low alcohol, and perhaps too smooth. No complex flavors.</li>
Belvino Prosecco Extra Dry</em>. (5/10). Cheap and just OK. Good for mixing in a spritz.</li> </ul> Orange</a></h4>

2023 Suisun Valley Assyrtiko, Hydrangea Skin-Contact</em>. (8/10). Excellent.</li>
Pielihueso Naranjo 2024</em>. (9/10). Very very good. Smooth, slightly fruity, and crisp. Best served cold.</li>
2023 Gulp Hablo Orange (1L)</em>. TBD.</li> </ul> Meads</a></h4>

The Remeadies</a> meadery near Sequoia National Park is very good. Go there if nearby. Unfortunately, their meads aren’t easily available in NorCal.</li>
Chaucer’s Mead</em>. (6/10). Very vanilla alcoholic honey flavor, smooth. Nothing objectionable, but nothing special either. Good for the price.</li>
Sky River Sweet Mead</em>. (7/10). At $30 it is expensive for what it is. Much more complex mead with a barnyard / wet grass flavor and floral notes. Not appreciated by everyone.</li>
Carroll’s Mead</em>. (7/10). A bit more honeyed vs Chaucer’s, but still relatively cheap. Smooth and tasty.</li> </ul> Sake</a></h3>
Takara Sake</a> has a sake factory in Berkeley, CA! I highly recommend visiting them and doing a tasting. What follows are all sakes from Takara (brewed in Berkeley).</p>

Sho Chiku Bai REI Junmai Daiginjo</em>. (8/10). One of our favorites. Cheap and good, serve chilled, crisp, fruity, and smooth. If it is too old, it will acquire a mushroom-y flavor.</li>
Sho Chiku Bai Organic Nama</em>. (6/10). Good chilled, but has an earthy flavor. Not suitable for a general audience.</li>
Sho Chiku Bai Takara Sierra Cold</em>. (7/10). This is an insane sake. Serve chilled. 12% ABV but tastes like zero! Very light flavor of slight sweetness and fruit, but mostly tasteless.</li>
Sho Chiku Bai Nigori Crème de Sake</em>. (9/10). Cream sake! One of our favorites. Serve chilled. Flavor is both fruity and earthy, more on the sweet side.</li>
Sho Chiku Bai Premium Ginjo</em>. (6/10). An acquired taste. Earthy, mushroom-y, not suitable for everyone.</li>
Sho Chiku Bai Mio Sparkling Sake</em>. (10/10). Low 5% ABV, crisp flavor, fruity. Everyone will love this.</li> </ul> Cocktail Liquors</a></h3>

Cointreau</em>. It works, I have no reference point for orange liqueur.</li>
Grand Marnier</em>. It works, great in a sidecar.</li>
Baileys</em>. Sweet, smooth, milky. Nothing bad, but nothing great either.</li>
Mr. Black Cold Brew Coffee Liqueur</em>. Decent, good espresso flavor.</li>
TJ’s Limoncello</em>. (8/10). Excellent, great in a lemon drop.</li> </ul> Amaro + Fortified Wines</a></h3>

Dolin Rouge (Sweet) Vermouth</em>. (8/10). Decent, I have no comparison point. Great to drink on its own too.</li>
Campari</em>. (7/10). Hard for a beginner to get used to its bitterness. Ease into it with ample dilution and sweeteners.</li>
Aperol</em>. (9/10). Works great in a spritz and really boosts appetite before a meal.</li>
Amaro Montenegro</em>. (10/10). The best thing I’ve ever tasted. Notes of vanilla and cola. Excellent in a spritz or just mixed with club soda (2oz amaro + 3oz soda water). This is a must buy.</li> </ul> Garnishes</a></h3>

Luxadro cherries</em>. TBD. These are so expensive, it better be worth it.
Totally worth it. So delicious and the syrup is great to add to a cocktail.</li> </ul> </li>
Angostura bitters</em>. It is what it is, pretty good — not many other options.</li> </ul> Equipment</a></h2>
There is some basic bar equipment that you need that I wish I had from the start. Just buy these things.</p>

Large ice cube tray</em>: I liked these Tovolo “King Cube” silicon trays (Amazon link</a>)</li>
Regular ice cube tray</em>: I bought some “no-spill” trays from OXO (Amazon link</a>)</li>
Jigger</em>: There are 2 major varieties (double-ended and angled). You need to be able to measure volumes from 1/4oz to 2oz. I think the angled jiggers are way easier to use at home with less spillage than the double-ended kind. I like this one from OXO (Amazon link</a>).</li>
Rocks Glasses</em>: Of course you need glasses to serve cocktails in. Of all the varieties, I find rocks glasses</a> to be the most versatile, albeit not the prettiest. I bought these ones from Anchor Hocking (Amazon link</a>). Coupes</a> or nick and noras</a> would be good next additions. If you want to splurge, RIEDEL</a> glasses are the best.</li>
Citrus juicer</em>: I recommend a manual wood citrus reamer — I got one from OXO (Amazon link</a>). Bulk juice citrus into a measuring jar and pour into your jigger. I no longer</em> recommend the NorPro 523 Stainless Steel Juicer (Amazon link</a>) — it doesn’t squeeze the citrus completely and leaves plenty of juice behind. For longevity and to avoid aluminum flaking, I lean towards a solid stainless steel juicer. I want to try the Barfly Large Manual Juicer sometime (Amazon link</a>).</li>
Mixing tin</em>: There are also cocktail mixing glasses, but I prefer a stainless steel mixing vessel. I have this one from American Metalcraft (Amazon link</a>).</li>
Hawthorne Strainer</em>: This OXO one is good (Amazon link</a>)</li>
Cobbler shaker</em>: Bartenders usually use Boston shakers, but a cobbler shaker is far easier to use at home. I got this one from OXO (Amazon link</a>). If I was OK spending $60 on this, I would buy the YETI one</a>.</li>
Fine mesh strainer</em>: I have a 3“ strainer from OXO (Amazon link</a>)</li>
Bar spoon</em>: Just get some cheap bar spoons, the exact model doesn’t matter.</li>
Corkscrew</em>: You need to be able to pull corks. Buy whatever is within your budget — these are all kind of the same. I have an OXO SteeL Winged Corkscrew (Amazon link</a>)</li>
Toothpicks</em>: The poor man’s version of the fancy cocktail pick — used to pick up cherries placed in the glass. Unfortunately, they are often too short to place over the rim of a glass.
Cocktail picks</em>: A substantial upgrade. Get stainless steel picks.</li> </ul> </li>
Cooler</em>: I use an insulated bag to hold all my ice and cold liquors, but a proper hard-case insulated cooler would be better.</li> </ul> My Shelf</a></h2>
I keep the fortified wines, meads, and cream liqueurs in my fridge, and the rest in a cabinet.</p>

Wine</dt>

Otima 10-Year Tawny Port</li>
Chateau Pey-Bonhomme Le-Tours</li>
Folk Machine Pinot Noir</li>
Belvino Prosecco Extra Dry</li> </ul> </dd>
Mead</dt>

Chaucer’s Mead</li>
Oliver Winery Camelot Mead Orange Blossom</li> </ul> </dd>
Whiskey</dt>

Four Roses Small Batch</li>
Mitcher’s US*1</li>
Redbreast 12 Year</li> </ul> </dd>
Rum</dt>

Diplomatico Reserva Exclusiva</li>
Appleton Estate 8</li>
El Dorado 3</li>
Probitas</li>
Clement Rhum Agricole</li>
Planteray 3 Stars</li> </ul> </dd>
Cognac</dt>

Kirkland XO Cognac</li> </ul> </dd>
Sake</dt>

Hakutsuru Sayuri Nigori Sake</li> </ul> </dd>
Soju</dt>

Oak 43</li> </ul> </dd>
Gin</dt>

Hendricks</li>
Citadelle</li>
Barr Hill</li> </ul> </dd>
Orange Liqueurs</dt>

Cointreau</li>
Grand Marnier</li> </ul> </dd>
Amaro / Fortified Wines</dt>

Campari</li>
Aperol</li>
Amaro Montenegro</li>
Dolin Vermouth Rouge (Sweet)</li>
Dolin Vermouth (Blanc)</li>
Dolin Vermouth (Dry)</li>
Sole Vermouth (Extra Dry)</li> </ul> </dd>
Other Liqueurs</dt>

TJ’s Sicilian Limoncello (x2) (stored in the freezer)</li>
Absolut Vanilla Vodka</li>
Mr. Black Cold Brew Coffee Liqueur</li>
Baileys Original Irish Cream</li> </ul> </dd>
Extras</dt>

Bitters</em>: Angostura, Orange</li>
Syrups</em>: Simple, Demerara, Grenadine, Orgeat</li>
Luxardo cherries</li>
Limes / lemons</li>
Heavy cream</li>
Chocolate syrup</li> </ul> </dd>
Books</dt>

Cocktail Codex</li> </ul> </dd> </dl>
To Buy</a></h2>
Bottles I want to try. Don’t look!</p>

Whiskey</dt>

Four Roses Single Barrel</li>
Larceny Small Batch</li>
Angel’s Envy</li>
Rittenhouse 100</li>
Nikka Whisky From The Barrel</li> </ul> </dd>
White Rum</dt>

Probitas</li>
El Dorado 3</li>
Flor de Cana 4</li>
Diplomatico Planas</li>
Planteray 3 Stars</li> </ul> </dd>
Aged Rum</dt>

Appleton Estate 8 / 12</li>
El Dorado 8 / 12</li>
Smith and Cross</li>
Rhum Barbancourt</li>
Santa Teresa 1796</li>
Diplomático Mantuano</li>
El Pasador de Oro Rum XO</li>
Kaniche XO Artisanal Rum</li> </ul> </dd>
Gin</dt>

Citadelle Gin</li>
Citadelle Jardin D’Ete Gin</li>
Tanqueray No. 10</li>
Barr Hill</li>
Monkey 47</li> </ul> </dd>
Cognac</dt>

Hennessy VSOP</li>
Pierre Ferrand Ambre / Original Formula</li> </ul> </dd>
Orange liquors</dt>

Rum-based curacao (dry): Pierre Ferrand Dry Curacao</strong>, Santa Teresa Rhum Orange</li>
Brandy-based curacao (sweet): Grand Marnier</li>
Triple secs (neutral): Cointreau</li> </ul> </dd>
Absinthe</dt>

St George Absinthe Verte</li>
La Clandestine Absinthe Superieure Liqueur</li> </ul> </dd>
Chartreuse</dt>

Chartreuse Green Liqueur (very difficult to find nowadays, can possibly find at K&L Wine Merchants)</li>
Chartreuse Vegetal de Grande Liqueur (a higher proof alternative)</li>
Dolin Genepy Le Chamois Liqueur (easier to find, combine with Chartreuse Vegetal to make a Last Word)</li>
Izarra Verte Basque Liqueur (Green Chartreuse “substitute”)</li>
Faccia Brutto Centerbe (another potential Chartreuse “substitute”)</li>
L’Excuse Herbal Liqueur (yet another potential “substitute”)</li> </ul> </dd>
Amaro</dt>

Aperol</li>
Amaro Nonino</li>
Amaro Montenegro</li>
Atost California Aperitif</li>
Cynar</li>
Campari</li>
Cappelletti Aperitivo</li>
Luxardo Aperitivo</li>
Ramazzotti</li>
Suze</li>
St George Bruto Americano</li>
Lillet Blanc</li>
Cocchi Americano</li>
Benedictine</li>
Grappa wine</li>
Ouzo</li>
Mastiha</li> </ul> </dd>
Extras</dt>

Raspberry syrup</li>
Castelvetrano olives</li>
Luxardo Maraschino Liqueur (essential to make a Last Word)</li> </ul> </dd>
Books</dt>

The NoMad Cocktail Book</li>
Death & Co: Modern Classic Cocktails</li> </ul> </dd> </dl>

All Hail the Magic Conch!

Unknown — Sat, 27 Sep 2025 00:00:00 +0000

Prelude</a></h2>
Deep in the depths of the ocean, in Bikini Bottom, there was one Club Spongebob 1</a></sup>. High up on a tall tree, sits the clubhouse, occupied by a sponge and Patrick.</p>
^1</sup>
Season 3, Episode 42a</a>. Watch the episode please.</p> </div>

</a> </div> </figure>
On this fine day, Squidward exits his Moai</a> to see these two barnacle heads laughing it up in their treehouse. Of course, there is nothing more irritating than their giggling. And when Squidward was told he wouldn’t “fit in” their club, he couldn’t take it. He had to join! And so up he went.</p>

</a> </a> </div> </div> </figure>
But once inside, he found out that, indeed, he couldn’t physically “fit in” the clubhouse and was now squished next to Bob and Patrick.</p>

</a> </a> </div> </div> </figure>
Squidward, desperate to escape, reaches for a branch and tries to pull himself down.</p>

</a> </a> </div> </div> </figure>
As he’s almost down, the branch rips apart, and the treehouse is launched into the sky (sea). The clubhouse flies through the sea and comes crashing down in the thick Kelp Forest.</p>

</a> </a> </a> </div> </div> </figure>
Oh no, Squid is trapped. Lost. Hundreds of miles from civilization.</p>

</a> </a> </div> </div> </figure>
He starts to panic. How can he escape these two morons now?</p>
But luckily, Squidward is part of Club Spongebob, and they have one trick up their sleeve: the Magic Conch</em>.</p>
</a> </div>
The Magic Conch is here to save the day!</p> </figcaption> </figure>

Squidward</strong>: You’ve got to be kidding! That is just a stupid toy! How can that possibly help us?!</p>
SpongeBob</strong>: [gasps] Squidward! We must never question the wisdom of the Magic Conch! The club always takes its advice before we do anything!</p>
Patrick</strong>: The shell knows all!</p> </blockquote>

</a> </a> </div> </div> </figure>

SpongeBob</strong>: Oh, Magic Conch Shell. What do we need to do to get out of the Kelp Forest?</p>
Magic Conch</strong>: Nothing</em>.</p>
Patrick</strong>: The shell has spoken!</p>
Squidward</strong>: Nothing</em>?! We can’t just sit here and do nothing!</p>
Squidward</strong>: I can’t believe you two are gonna take advice from a toy!</p> </blockquote>

</a> </a> </div> </div>
Just do what the Conch tells you to do</p> </figcaption> </figure>
Squidward thinks he’s smarter than the Magic Conch, so he runs and runs to escape the forest, only to find out he’s running in circles and is right back at Club Spongebob. He sets up a camp with the resources he has while SpongeBob and Patrick continue to listen to the Conch.</p>

</a> </a> </div> </div>
Squidward tries to use his own thinking rather than obeying the Conch</p> </figcaption> </figure>
Right as Squidward was about to enjoy his roasted sea insect, a miracle falls from the sky, right into the camp of SpongeBob and Patrick! What a gift from the Magic Conch!</p>

</a> </a> </div> </div>
The gift from the Conch has arrived!</p> </figcaption> </figure>
As SpongeBob and Patrick enjoy their feast, Squidward begs to be allowed to touch the food. They inform Squid that only the Conch</em> can approve his request. And so he asks the Conch, again and again, “may I have something to eat?”</p>

</a> </a> </div> </div> </figure>
But he didn’t give the Conch its due earlier, so the Conch only gives him a simple “no”. Squidward goes crazy as the Conch seems to be aware of his dismissal of its powers.</p>
And then, someone cuts a path into the forest!</p>

</a> </a> </div> </div>
The Kelp Forest ranger has arrived to save Squidward!</p> </figcaption> </figure>
Squidward screams in delight — someone is here to save them! But to Squid’s dismay, the ranger also has a Conch of his own.</p>

</a> </a> </div> </div> </figure>

Kelp Forest Ranger</strong>: All right, Magic Conch… what do we do now?</p>
Magic Conch</strong>: Nothing</em>.</p>
SpongeBob, Patrick, and Kelp Forest Ranger</strong>: All hail the Magic Conch!</p> </blockquote>

</a> </a> </div> </div> </figure>
Sound Familiar?</a></h3>
And so they sit, in silence, in devotion to the Conch, until another miracle falls from the sky. Perhaps, if they ask the Conch again and again, it will give them the answer to their problem. Perhaps, a new version of the Conch will do better than the old one.</p>
Whenever you have a question, don’t think, just ask the Conch. And whatever the Conch tells you, is what you should parrot to others and follow diligently.2</a></sup></p>
^{2</sup>
The Conch = a large language model (henceforth called “the model”)</p>
</div>
But, just because it worked for SpongeBob, doesn’t mean it will work for you.
There is a growing mass of people who believe the Conch will enable them to do nothing, and that good things will fall out of the sky.</p>}
The Situation</a></h2>
There has been a lot said over the past two decades about how electronics can sap away our thinking powers. When humans lose the ability to be bored</em> and produce something from nothing, their mental faculties decay. I believe Cal Newport put this phenomena into the public consciousness with his book “Deep Work”3</a></sup> where he discusses how electronics, especially phones, have put humans in an unprecedented situation where they can go through life and never</em> be bored.</p>
^3</sup>I started hearing about the importance of boredom from 2016 or so when Newport’s book was published. A counterpoint can be found here</a>.</p> </div>
In the past two years, in addition to the phone, which can prevent even a few seconds of boredom from setting in, we now have “the model”. Whatever attention span degradation phones have caused, has been or will soon be, dwarfed by the advent of the model.</p>
As people begin to use the model, they start by asking it a few questions about topics that they would have used a search engine for in the past. However, as model dependency grows, people begin to outsource their thinking</em> and even *thoughts</em> wholesale to the model. I need not dwell on this point too much, since others have made it much better.**4</a></sup></p>*
^4</sup>
See the article “The End of Thinking”</a> by Derek Thompson</p> </div>
Bimodal Undergrads</a></h3>
I’ve been in the university system for a long time. Almost certainly too long. Every year, I get a sample of the current class of undergrads to examine — both from teaching classes and from advising them on research. Year after year, the mean quality of undergrads has degraded — the average undergrad is increasingly motivated by money, status, and job prospects rather than any intrinsic interest in computer science. However, this trend about averages says nothing about the extremes.</p>
The mediocre undergrads are crashing to the level of the barely literate ones, while the elite undergrads can rival senior grad students in programming competence, inquisitiveness, and instinct. Just in the past year, I’ve seen a few truly spectacular undergrads the likes of which I haven’t seen before. They can work autonomously, ask all the right questions, learn on the fly, and somehow have enough time to be at the top of their classes and do self-driven research. I know that when I was an undergrad, I was very far below their level.</p>
You can probably predict why this bimodal distribution has emerged and gotten more extreme recently. It’s the model.</p>
What’s happening is that even the 80th percentile undergrads are falling victim to model dependency. The majority of undergrads are not only preempting their boredom via their phones, but they are also preempting their thoughts via the model. We often see undergrads that produce reams of code from the model, but are unable to explain what is going on, and more importantly, what are they trying to do.</p>
The elite undergrads are always in control of their own thoughts. They use the model as just an enhanced search engine, which gives them confidence to enter new areas. Their knowledge compounds rapidly as they use the model to pull information, but use their own brains to synthesize it.</p>
Thinking in “Embedding Space”</a></h3>
After talking about undergrads, I must now talk about those ‘above’ me: the professor class</em>. If you thought professors were speaking in meaningless platitudes before, you haven’t seen anything.5</a></sup> I swear, if I could peer inside the head of a typical professor, I’m sure all I would find is a vector database.</p>
^{5</sup>
There are still some intelligent professors, but among the mediocre, the recent degradation has been extreme</p>
</div>
I often joke that professors (and the model) think in “embedding space”.
What I mean to say, is that they think as if embeddings are semantics</em>.
I can’t blame them. After all, if you ask the model, it will claim that embeddings capture semantics</em>.</p>
When they speak, they will put words together that appear</em> close in the embedding space, but are actually unrelated with respect to their true semantics.
This is just a matter of training data: both for the professor and the model.
Without enough good quality data, embeddings will tend to just capture which words occur together, rather than how they are related.
While a professor should be able to ground word semantics in reality, the model cannot, so I can give the model a pass here.</p>}
Biological Frontends for Mr. Model</a></h3>
What could be worse than people becoming dependent on the model? What if people became</em> the model? Indeed, this is what I witness nowadays. Professors have become biological frontends for the model.6</a></sup></p>
^{6</sup>
I don’t mean to bash on professors too much. All professions have degraded thusly.</p>
</div>}
When a student asks for advice, the questions are forwarded to the model, and the model’s response comes out of the professor’s mouth. The professor asks the model for research ideas, and then recites its responses to their students. When reviewing papers for a conference, the PDFs are fed straight into ChatGPT Pro 7</a></sup>, and its outputs are massaged into the HotCRP boxes.</p>
^{7</sup>
It’s research-grade intelligence after all</p>
</div>}
Context Pollution</a></h4>
Models suffer from context pollution</em>, where garbage can accumulate in its context window that distracts the model from the task its supposed to perform. Being able to tell what the relevant context is, while discarding the rest, is an essential aspect of intelligence.</p>
If you look at the managerial class in general, they are increasingly falling victim to context pollution. They attend all kinds of useless meetings and conferences where the executive class jumbles together words that have no relation to each other. Their heads are filled with words from these continuous meetings and, when they are prompted to discuss something with their subordinates, they bring up unrelated nonsense from their context window. Buzzword speak has become more and more ubiquitous.</p>
“Asking the Model” as Research</a></h2>
Day after day, I will look at my Google Scholar notifications and will see the same paper repeated 100 times. Here are some recent examples:</p>

VerilogMonkey: Exploring Parallel Scaling for Automated Verilog Code Generation with LLMs</a></li>
Automated Multi-Agent Workflows for RTL Design</a></li>
QuArch: A Question-Answering Dataset for AI Agents in Computer Architecture</a></li>
AutoChip: Automating HDL Generation Using LLM Feedback</a></li>
Chip-Chat: Challenges and Opportunities in Conversational Hardware Design</a></li>
RTLCoder: Fully Open-Source and Efficient LLM-Assisted RTL Code Generation Technique</a></li>
Multi-Agent Reinforcement Learning for Microprocessor Design Space Exploration</a></li>
Large Language Monkeys: Scaling Inference Compute with Repeated Sampling</a></li>
CodeMonkeys: Monkey SWE, Monkey Do</a></li> </ul>
All of these thousands of papers just amount to “asking the model” in a loop, with some scaffolding, some tool use, some “prompt engineering”, and perhaps some training data for supervised fine-tuning. Perhaps some will conduct a beam search across many samples provided to them by Mr. Model. Perhaps some will construct a new dataset from thin air, which they pose as a “benchmark”.</p>
Of course, this LLM mania isn’t limited to computer architecture papers. We see this in all disciplines, where the ‘solution’ proposed by a paper is “asking Mr. Model”, and the ‘problem’ is whatever fake problem they invent.</p>
Furthermore, due to the high latency of conference paper deadlines and review cycles 8</a></sup>, by the time a paper is published, its “results” are already suspect. The authors will have run all their experiments with Mr. Model v7 and by the time the world sees their work, Mr. Model v8 has been released, and large parts of their specialized ‘prompting’ techniques and scaffolding are made obsolete.</p>
^{8</sup>
It will usually take 6+ months from running experiments to public paper release</p>
</div>}
It is just too easy for a professor to latch onto the “ask the model” methodology and enjoy the feeling of being part of the hype. At this point, “asking the model” has become its own area of research. Why bother using the model as a productivity booster or powerful search tool to produce impactful research, faster 9</a></sup>, when you could just ask Mr. Model to write a bunch of kernels and report that?</p>
^{9</sup>
You will find that Mr. Model becomes useless very quickly when exploring an untouched area</p>
</div>}
My final point is that academics should work on projects that have substantial intellectual, logistical, and financial risks; risks that are so high that industry researchers would not take on such projects. What is the risk here? That Mr. Model may not always be correct? Why are academics “asking the model” when there are VCs investing billions of dollars and accepting infinite risk for startups to pursue “LLM for X”?10</a></sup></p>
^{10</sup>
I can appreciate that some academics use papers as a launching pad for a startup. That’s fine I guess</p>
</div>}
“Asking the Model” as a Course</a></h2>
Not only has “asking the model” become its own area of research within an engineering domain (e.g. RTL design, GPU performance engineering), but “asking the model” more generally has become an acceptable topic for courses. Let me go over three examples from Stanford, Berkeley, and Harvard.</p>
Stanford</a></h3>
CS329A: Self-Improving AI Agents</a> is being offered for the second time at Stanford.</p>

The course will start with self-improvement techniques for LLMs, such as constitutional AI, using verifiers, scaling test-time compute, combining search with LLMs, and train time scaling with RL. We will then discuss the latest research in augmenting LLMs with tool use, code, and memory, and orchestrating AI capabilities with multimodal interaction. We will next discuss multi-step reasoning and planning problems for agentic workflows, and the challenges in building robust evaluation frameworks.</p> </blockquote>
It all sounds very fancy, but it boils down to asking the model in a loop. I believe this paper (The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery</a>) is emblematic of the method being promoted. If this method can produce ‘papers’ that are so ‘high-quality’ that even humans would accept them to a ML conference, then we should be questioning the value of what humans are working on.</p>
With that being said, the reading list for this class isn’t bad at all. These papers are interesting and worth skimming just to understand the state of the field. All I ask is that the students don’t make their class projects about getting Mr. Model to solve some problem.</p>
UC Berkeley</a></h3>
CS294: Disrupting Systems Research with AI</a> is being offered for the first time at UC Berkeley.</p>

We are now at the beginning of a significant shift, where a new class of AI tools can autonomously generate algorithms that match and sometimes exceed the best human-designed solutions.</p> </blockquote>

This course explores the frontiers of this new methodology, examining the future role of the researcher as a “strategic advisor” who guides powerful AI assistants rather than manually engineering solutions.</p> </blockquote>
This is quite similar to the Stanford class, albeit a bit more hyped up. The proposed methodology boils down to “vibe research”. Ask the model to propose grand changes to some existing codebase. Then, ask the model to produce some code to manipulate a repo and see what happens. Just go with the flow.</p>
“The shell knows all! The shell has spoken!” — Patrick Star</p>
Harvard</a></h3>
And now, this is the most egregious example of model-brained nonsense by far.</p>
Presenting, CS249r: Architecture 2.0</a>, being offered at Harvard for the first time. I urge the reader to read the “blog posts” on the website, such as: “Week 2: The Fundamental Challenges Nobody Talks About”</a>. It is very obvious that all these “blog posts” are written by Mr. Model. In fact, the entire website</em> is generated by Mr. Model.</p>
</a> </div>
I couldn’t have said it any better myself. See the class slides</a>.</p> </figcaption> </figure>
This class is about a concept that Prof. Vijay Janapa Reddi has coined: “Architecture 2.0</a>”. All this boils down to is generating tons of “data”11</a></sup>, training some models on that data, and hoping for the best. After all, if this method has yielded good results in image classification and English emission, then surely the same method will yield fruit in computer architecture. It’s just a matter of more and more data, and more and more compute: this is the “Bitter Lesson” at work.</p>
^{11</sup>
“Data” includes random RTL designs, gate-level netlists, PPA estimates, instruction traces and so forth: the “corpus” of computer architecture.</p>
</div>}
Astute readers will note that Reddi’s “Architecture 2.0” is just a ripoff of Andrej Karpathy’s “Software 2.0”</a>. But recently, at YC’s “AI Startup School”</a>, Karpathy presented his talk, Software Is Changing (Again)</a>, where he coined “Software 3.0”!</p>

</a> </a> </div> </div>
Karpathy’s slides on Software 3.0</p> </figcaption> </figure>
So, Prof. Reddi, Architecture 2.0 has already been superseded! It should be time for Architecture 3.0: aka “ask the model”.</p>

</a> </div>
Model-generated text that contrasts Architecture 1.0 with 2.0 (actually Architecture 3.0)</p> </figcaption> </figure>

</a> </div>
Architecture 3.0 methodology in practice. Ask the model (in a loop).</p> </figcaption> </figure>
I have collected many model-isms from Prof. Reddi’s course and placed them in the addendum of this article</a>.</p>
A Better Path</a></h2>
As was discussed before, the focus should not be on asking the model and evaluating how it does, but rather on using the model to build better software. I recently saw a good example of this from Mark Ren’s group 12</a></sup> called SALUTION: Autonomous SAT Solver Evolution</a>. They aren’t presenting “asking the model” as the goal itself, but instead asking the model to produce a better SAT solver, which is the main contribution. They can discuss the specific elements of the code</em> that the model generated to produce a faster solver. The focus is on the product, not the model.</p>
^{12</sup>
Mark Ren and his team at NVIDIA Research are one of the few competent actors in the field of “ML for CAD”</p>
</div>}
Our new framework, SATLUTION, autonomously evolves Boolean Satisfiability (SAT) solvers via LLM agents that outperformed the 2025 SAT Competition champions by more than 10%</p> </blockquote>
It would be good if more work like this was done. Work as a domain expert who is trying to improve some algorithm or work on a specific problem and use the model (who cares how you’re using it) to help you make things better.</p>
Another recent example is the paper “ASPEN: LLM-Guided E-Graph Rewriting for RTL Datapath Optimization”</a>. Here, e-graphs are used to perform word-level datapath optimization, which is usually done by hand for arithmetic circuits. The LLM is used to generate additional rewrite rules for the e-graph engine and prioritize them based on PPA feedback from the CAD tools. This process, which would normally require human intervention, is done autonomously, and it produces improved results over a static rewrite rule set. While I believe it would be better to fold back these improvements into more robust abstractions in the e-graph engine, I can appreciate that the model did do something cool.</p>
Conclusion</a></h2>
Just as one should hesitate when picking up their phone to mitigate a moment of boredom, one must hesitate before shooting a query at the model. Think! Losing your boredom is bad enough, but losing your sovereignty is even worse. There is a huge risk that biological general intelligence will dry up way before AGI can come on the scene to save us.</p>
But, you know what? Perhaps I’m wrong. Perhaps “asking the model” is the most useful, impactful, and important thing we can all do today.</p>
In the next few years, I may be the one saying:</p>

Squidward</strong>: All hail the Magic Conch</em></p> </blockquote>
</a> </div> </figure>
Response to Objections</a></h2>
I appreciate criticism of anything I wrote in this article. Just email me.</p>
Let me begin by saying this article is not</em> a formal academic argument, as it should be clear. It is just a rant. However, if I were to distill the article into some statements I believe in, here are a few:</p>

Do not use the model as a substitute for thinking or learning. You want to be a domain expert. You want the ability to produce insightful thoughts. Use the model as a search and exploration tool to get to that point first. Only then, can you can begin to use the model as the creator / editor of a codebase (or a piece of writing, or so forth).</p> </li>

Academics should be careful about what they work on to avoid getting on a hype train they aren’t capable of riding. Academics aren’t rich, both in terms of free cash and GPU capital. The risks that academics take should be wholly different than those taken by VCs or megacap corporations.</p> </li>

The model is powerful. It’s getting more powerful every month. You would be a fool to dismiss the power of the model and not use it altogether.</p> </li> </ol>
Now, I’ll discuss some objections to this article and give my responses.</p>
Agents are Hard</a></h3>

Agents are not as trivial as you make them out to be. They aren’t just a combination of prompts, RAG, fine-tuning, tool calling, tool output post-processing, beam search, hardware / profiler feedback, and so forth, in a big loop.</p>
If you look at SOTA agents like Cursor, Claude Code, Codex CLI, Cline, Windsurf, Aider, Devin, and so forth, they are growing in complexity, rather than their scaffolding being subsumed by the base model. In fact, as the models become more powerful, and their tool calling abilities improve, the scaffolding may have to become more complex to accommodate these new abilities.</p> </blockquote>
I agree that reliable agents are hard to build, and I do not want to take away from the valuable engineering and research efforts invested in these agents. However, I believe that to do good work in the area of agent architecture / optimization / sampling, one needs to have real use cases and evaluation criteria, beyond just the published “standard” benchmarks (e.g. SWE-Bench). Furthermore, agent architecture research is very expensive when it comes to running evals for each proposed change.</p>
All of this is to say, that if one wants to pursue research in this area, they should either (1) go to Anthropic, Cognition, or the like and work on real problems or (2) take the VC money that is freely flowing and build a domain-specific agent or (3) contribute to open source agents such as Aider.</p>
I believe it isn’t reasonable to do coding agent work in academia anymore — it was reasonable for a short period of a year or so, but today the resources needed to make a dent are too substantial. Academics are flooding this area because (in my opinion), it feels easy to pick a niche domain and build an agent for it that outperforms some stale baseline on a niche benchmark. I have to say though that I’m a fan of DSPy</a> and I think work along this line should be pursued by academics.</p>
“Asking the Model” Does Work</a></h3>

As the readings from these classes have shown, asking the model does indeed work to optimize existing software or even propose new research ideas, produce evaluations, and write an entire paper that passes scrutiny by humans. You pointed out research on optimizing SAT solvers using LLMs — this is another example of the model being able to produce useful outputs given the right scaffolding.</p>
How can you claim that “asking the model” should be avoided, when we can see that is does yield fruit? How can you claim that academics shouldn’t pursue agent engineering research when it’s about how to leverage LLMs effectively to improve software / systems?</p> </blockquote>
Firstly, I agree that the “asking the model” does often produce good results, and that the right scaffolding around the model greatly impacts the quality of its results. However, I do believe that some discretion is needed here with regards to choosing the right problem to work on and leveraging existing agentic frameworks vs rolling your own. My point is that the result of domain-specific research should be the product rather than the model loop that is used to create the product.</p>
Furthermore, it is much more valuable to build something using the model and report how it went via a blog post rather than evaluating some bespoke agent you built on some benchmark and publishing the results for them to only arrive 6 months in the future. Here is a good example: Vibing a Non-Trivial Ghostty Feature</a>. More examples can be found on Simon Willison’s blog</a> such as “Getting DeepSeek-OCR working on an NVIDIA Spark via brute force using Claude Code”</a>.</p>
Lastly, I believe that domain knowledge is crucial when it comes to building the right scaffolding, choosing a good problem, and picking a strong baseline to compare against. Building that domain knowledge should be the priority for anyone entering a domain. Asking the model without understanding the domain is vibe research, and should be avoided.</p>
The Magic Conch is not an LLM</a></h3>

Your analogy is bad. The magic conch is similar to a magic 8 ball: it can’t give sophisticated answers to questions, but rather just picks random phrases from a pre-programmed list. The Conch doesn’t consider the context of the question you ask it, it doesn’t possess a world model, and it cannot use tools to prepare its answer.</p>
In contrast, Mr. Model is a very powerful statistical token generator trained on a massive corpus of world knowledge that has common sense and can reason. How can you claim these two things are equivalent, when they are so obviously not?</p> </blockquote>
Mr. Model is certainly more sophisticated than the Conch, there can be no doubt. However, the comparison has more to do with how the Conch is used rather than what it says. Uncritically taking the Conch at its word, following its instruction as if it came from the Gods, and delegating every question you have immediately to the Conch are the problems. Finally, for academic researchers, leveraging</em> the Conch to do interesting things is far more valuable than sweeping the space of using</em> the Conch.</p>
Don’t take this analogy too seriously.</p>
Barbarians at the Gate</a></h2>
Shortly after I published this article, a whitepaper was published on arXiv on October 7th titled “Barbarians at the Gate: How AI is Upending Systems Research”</a> by the Berkeley Sky Lab</a> professors (Koushik Sen, Matei Zaharia, Ion Stoica, …). This whitepaper argues that “asking the model” works, is effective, and should be the basis for what they term “AI-Driven Research for Systems</em>”.</p>

Given a task, the typical AI-driven approach is (i) to generate a set of diverse solutions, and then (ii) to verify these solutions and select one that solves the problem.</p>
We argue that systems research, long focused on designing and evaluating new performance-oriented algorithms, is particularly well-suited for AI-driven solution discovery.</p>
We term this approach as AI-Driven Research for Systems</strong> (ADRS), which iteratively generates, evaluates, and refines solutions.</p>
We then discuss the broader implications for the systems community: as AI assumes a central role in algorithm design, we argue that human researchers will increasingly focus on problem formulation and strategic guidance.</p> </blockquote>
The traditional model of systems research involves finding some problem in an existing codebase and then hacking on it until you get a sufficiently good solution that warrants a paper.</p>
</a> </div>
The ‘typical’ methodology of systems research.</p> </figcaption> </figure>
The whitepaper proposes a way to automate the most time-consuming part of systems research: the process of editing code and tuning / sweeping / evaluating.</p>

</a> </div>
AI-Driven Research for Systems (ADRS) is, quite unambiguously, ‘ask the model’ (in a loop).</p> </figcaption> </figure>
ADRS involves asking powerful agents to edit / evaluate code iteratively using a genetic algorithm to guide the search process (ala AlphaEvolve</a>, OpenEvolve</a>, GEPA</a>).</p>
The argument the whitepaper uses to justify “ADRS” goes like this:</p>

Systems research is about finding a bottleneck or sub-optimal algorithm in a software/hardware stack (e.g. compilers, runtimes, schedulers, hardware tuning knobs) in order to improve some application’s performance</li>
The model is very good at editing code, performing search, and mixing algorithms/heuristics from the literature in a novel way. The code that the model needs to edit is often localized to a few files (e.g. app-level scheduler algorithm tweaks).</li>
It is easy to evaluate the impact of the model’s edits reliably and quantitatively with respect to correctness and performance characteristics (e.g. memory usage, latency, throughput). Evaluation can be done using performance simulators and/or real systems.</li>
The model is getting cheaper and is able to maintain coherence across longer time horizons than before. The cost to ask the model thousands of times to solve a problem is marginal</em> compared to the cost and time for a human to implement optimizations in a codebase.</li> </ol>
Therefore</strong>: Asking the model is a suitable and effective way of conducting systems research. Humans should focus on guiding</em> the agent and formulating</em> useful problems to solve.</p>
My Opinion</a></h3>
At first glance, everything here seems reasonable. I would say that their argument is airtight, except for the first premise</em>.</p>
What is Being Argued For?</a></h4>
Here is the sentence from the whitepaper that I paraphrased into the first premise:</p>

In this paper, we advocate an AI-driven approach to systems performance problems. While performance optimization is not the only focus of systems research, it remains a central one—a brief survey of top systems, networking, and database venues (NSDI, OSDI, SIGMOD, SOSP, and VLDB) shows that over one-third of published papers feature performance optimization algorithms as their core contribution.</p> </blockquote>
The authors concede that not all of systems research can be boiled down to performance optimization. However, for the sake of argument, and since ADRS is most suitable for it, let’s first examine the case of performance optimization systems research.</p>
In this case, I would argue that the paper has already been written</em> before any work has been done. The researcher has found a problem (i.e. bottleneck, performance pathology) in an existing system that he knows is solvable (i.e. amenable to optimization) using the typical techniques used in the field (e.g. better heuristics, more suitable data structures, speculation). It is just13</a></sup> a matter of doing the profiling and implementation to show that an improvement has been made and to characterize how significant it is.</p>
^{13</sup>
I don’t mean to trivialize the difficulty of implementation, but it is often mechanical in nature (which is why ADRS can work)</p>
</div>}
In the case of ADRS, the situation is even more extreme — the solution is known even before the problem has been specified. The solution being, “ask the model”.14</a></sup> Perhaps the details of prompting, test-time scaling, and Monte Carlo tree search aren’t known in advance, but this doesn’t change the fact that the model is primal</em> in ADRS.</p>
^{14</sup>
“Asking the model” is a solution in search of a problem, a ‘solution-oriented’ problem.</p>
</div>}Why Can’t We Go All the Way?</a></h4>
Consider this statement from the whitepaper:</p>

In the broader context of AI-driven research, our focus is deliberately narrow. Not only do we restrict our scope to the systems domain, but in this context, we focus only on the task of solution discovery, while largely ignoring other aspects in the research process, like problem formulation, literature survey, or paper writing.</p> </blockquote>
ADRS is focused on the task of writing / editing code (which they term “solution discovery”), while the other aspects of the research process (problem identification, setting up the model’s scaffold, writing a paper) are left in the hands of humans. But if we were to consult the charts</a>, it seems that Mr. Model can take over these other aspects too, very soon.</p>
Why can’t the last step (“Paper Write-Up”) be done agentically? This is often the most mechanical and uninspiring part of the research process. We can even “close the loop” by having the model come up with new problems to solve on its own — this is just a matter of search over existing code and literature + profiling production machines.</p>
The argument for ADRS, taken to its logical conclusion, would end up turning systems research into a fully model-driven paper pumping machine. Arguably, systems research in its current form is already a human-driven paper pumping machine. We can already see the first inklings of ADRS transitioning the paper machine from being human to model driven.15</a></sup></p>
^15</sup>
This methodology has been put into action by Sakana AI</a></p> </div>
What Should Academics Do?</a></h4>
Even if we were to concede that ADRS is a good research methodology, what exactly does this process produce? In the best case, it produces a system (e.g. an application server, inference service, load balancer, workload scheduler) that is slightly more optimized than what already exists.</p>
This form of research is justifiable</em> and undoubtedly valuable</em> for a large hyperscaler such as Google: see their paper “ECO: An LLM-Driven Efficient Code Optimizer for Warehouse Scale Computers”</a>. They show the model many historical examples of performance anti-patterns and their fixes, and then ask the model to find additional opportunities in the `google3</code> codebase. The model autonomously makes the right code edits, checks that its changes are valid with existing unit/integration tests, makes a PR for a Googler to review, and once deployed in production, the model-written code is monitored to validate its correctness and characterize its performance improvement.</p>`
Currently deployed on Google’s hyperscale production fleet, this system has driven >25k changed lines of production code, across over 6.4k submitted commits, with a >99.5% production success rate. Over the past year, ECO has consistently resulted in significant performance savings every quarter. On average, the savings produced per quarter are equivalent to over 500k normalized CPU cores.</p> </blockquote> How can anyone argue with this? Indeed, I must concede that this work is valuable and impactful.</p> While we should be impressed by this research, we should also ask, “is this what academics should be doing”</em>? It makes sense for a hyperscaler to optimize every bit of code they can, where even a tiny performance optimization makes a significant impact on the scale they operate at. But does it make sense for an academic research lab to do the same?</p> I would argue the answer is no</em>. For one, academics should be prototyping new systems</em> rather than making nit optimizations to existing ones. If academics are relegated to doing the same work as industry research labs, but on open source repositories, it begs the question as to why academics exist in the first place. Where is the risk in ADRS? Do something the industry players can’t do!</p>What is Systems Research?</a></h4> Let’s go back to the original point I made about the first premise being fishy. As the authors concede, systems research is not</em> just about finding some bottleneck and alleviating it with the right code edits (“performance optimization”). If ADRS picks up steam and begins to dominate the conference proceedings, which I suspect it will, it will crowd out the other, more interesting, aspects of systems research.</p> I claim that the most impactful and interesting systems research projects involve building completely new libraries, tools, abstractions, programming models, compiler infrastructures, and paradigms</em>. Let me illustrate by example.</p> Just from the lab of the ADRS authors, came these wildly successful projects: Mesos</a> (which led to its modern incarnation, Kubernetes</a>), Spark</a> (which pioneered modern data processing abstractions and was commercialized as Databricks</a>), Ray</a> (which is widely used for distributed AI jobs, arguably a successor to Spark, and was commercialized as Anyscale</a>), and Skypilot</a> (which implements a “cloud-neutal” abstraction layer and will certainly be commercialized soon).16</a></sup></p> ^{16</sup> Other impactful systems projects: ML (TensorFlow, PyTorch, JAX, MLIR, Triton), data (Dask, Polars, DuckDB), infra (gRPC, Terraform, Kafka, Nix), databases (Cassandra, RocksDB, ClickHouse)</p> </div> While these projects are arguably too industrial and commercialization-focused for academic research, their impact can’t be understated. Impactful systems research involves building real systems and new abstractions, and free-form exploration</em> of some unknown space. Can the ADRS methodology alone produce high impact academic research? It seems unlikely.</p>}Addendum</a></h2> With all the argumentation out of the way, it’s time for fun. I’ll use this section to write about some funny model-isms from Prof. Vijay Janapa Reddi</a>’s CS249r (Architecture 2.0: Agentic AI for Computer Systems Design</em>) website</a>.</p> Reddi’s Agents</a></h3> Reddi uses a “discussion agent”</a> that writes ‘discussion questions’ to be discussed in class.</p> </a> </div> </figure> I’m sure the students will appreciate an agentic discussion.</p> </a> </div> The students shouldn’t bring AI-generated responses to the AI-generated discussion</p> </figcaption> </figure> Reddi uses a “blog writer agent”</a> that provides “editorial feedback”.</p> </a> </div> Mr. Model, please ‘present the draft blog post to the instructor’</p> </figcaption> </figure> The “blog writer agent” needed some coaxing</a> to not produce text that was too easily attributable to Mr. Model.</p> </a> </div> Mr. Model, ‘AVOID LLM Writing Patterns</strong>’. It didn’t work 😆</p> </figcaption> </figure> The “blog writer agent” knew that Claude was the actual author of the model-generated prose, so Reddi had to patch that</a> too.</p> </a> </div> Mr. Model, don’t attribute yourself as the author!</p> </figcaption> </figure> When faced with a moral quandary about authorship, it is best to ask the model. The model’s opinion is on Reddi’s syllabus</a> of course.</p> </a> </div> Transparent attribution</strong>: When AI significantly contributes, acknowledge it. How about it Reddi?</p> </figcaption> </figure> I will remind the reader that Prof. Vijay Janapa Reddi is a tenured associate professor</em> at Harvard University. Undergraduates pay Harvard $59k in tuition every year to be “taught” by biological model frontends like Prof. Reddi.</p> The Model Goofs Up</a></h3> Look at the “Week 5” section</a> of the syllabus.</p> </a> </div> </figure> Near the bottom, do you see the paper that’s supposedly titled “Reinforcement Learning for FPGA Placement”? Why is that there? Isn’t this section supposed to be about “GPU Kernels and Parallel Programming”?</p> If you look at the linked paper</a> it’s actually titled: “Dynamic minimization of bi-kronecker functional decision diagrams”. So it seems Mr. Model just put some random paper in the reading list that seemed close in the embedding space.</p> The great irony, is that on the syllabus page</a>, the model has already warned about itself!</p> </a> </div> </figure> “Verify everything”, right? But it seems the professor himself is incapable of using GenAI effectively. The model is more aware than the people using it.</p> The Model Cleans Up Model-Generated Text</a></h3> See this commit</a> where Reddi uses Mr. Model to remove model-isms from the “blog posts”, such as em dashes and hyphenated words.</p> </a> </div> Mr. Model, please don’t use em dashes. It’s too embarrassing. Use colons instead.</p> </figcaption> </figure> In this commit</a>, Prof. Reddi uses Claude Code to “polish” a verbose “blog post” likely originally created by Gemini or GPT (or perhaps Claude without the “avoid LLM writing patterns” instruction).</p> </a> </a> </div> </div> Mr. Model ‘polishes’ its own text. Slop → slop.</p> </figcaption> </figure> The Model Goofs Up Again</a></h3> As the model was removing model-isms from its own text, it also made a mistake</a>!</p> </a> </div> Mr. Model goes schizo and edits a correct link for no reason</p> </figcaption> </figure> It inadvertently updated the link of the ECO paper</a> to a random paper</a> titled “Positivity Bounds on Massive Vectors”. I didn’t know that Reddi was so interested in high energy physics! The incorrect link to the ECO paper is still up on the Week 4 blog post</a>.</p> Thinking in Embedding Space</a></h3> Take a look at Prof. Reddi’s latest model-generated “blog post”: Week 5: From CPU Transparency to GPU Complexity - The Performance Engineering Frontier</a>.</p> There is one section that is quite revealing:</p> </a> </div> </figure> maintaining low latency in the prediction process is crucial for AI coding assistants</p> </blockquote> OK sure. Cursor has to optimize for low time-to-first-token to please its typical user when not operating in agent mode (e.g. tab completions, directed code edits, code analysis).</p> He noted that the cost of efficient inference scheduling, both for rollout generation and reward computation, often dominates the training pipeline. This observation captures a fundamental tension in production AI systems: the very optimizations that make training efficient can create bottlenecks during inference</p> </blockquote> So perhaps inference can dominate the total runtime of an RL pipeline. But why would any training optimizations affect the performance of the model when just running inference?</p> While research systems can afford to spend significant time optimizing kernels offline, production systems must balance optimization quality against response time.</p> </blockquote> What? Kernel optimization is almost by definition done offline. Why would a “production system” have to “balance” anything? Why is “response time” important when you’re optimizing a GPU kernel? Even if we were to concede that TTFT makes an agentic optimization loop run faster, it is still throughput-bound.</p> This is a great example of the model thinking in embedding space</em>. It first notices that “low latency” is important in interactive LLM applications. Then it jumbles up the meaning of the word “latency” in different contexts in the subsequent paragraphs and produces unintelligible slop that sounds like a deep insight.</p> Let’s Ask the Model</a></h4> Of course, I had to check if the model could understand what I mean by “thinking in embedding space”. After placing the right embeddings in its context window, it began to “understand”.</p> </a> </a> </div> </div> GPT5 does a good job explaining the concept of ‘thinking in embedding space’</p> </figcaption> </figure> Funnily enough, GPT kept suggesting that it could follow up with mocking quips about “thinking in embedding space” — perhaps it is too tuned to my style of conversation.</p> </a> </a> </div> </div> GPT5 is very eager to mock. AI safetyists must be concerned.</p> </figcaption> </figure> An Agentic Textbook</a></h3> Prof. Vijay Reddi has become quite confident in the ability of the model. Now, he is just asking Claude Code to rewrite huge swaths of his textbook “Machine Learning Systems: Principles and Practices of Engineering Artificially Intelligent Systems”</a>.</p> For instance, take a look at this commit</a> which had Claude Code rewrite thousands of lines of prose, all at once.</p> </a> </div> </figure> What is the chance he reviewed these changes personally, one-by-one? Consider that the next massive commit from Claude is pushed just an hour</a> after this one.</p> Of course, nothing is being reviewed. We are now also in the era of “vibe writing” — Mr. Model produces mountains of verbose and meaningless prose that goes into a textbook. The undergrads better get used to it.</p> Mr. Model, Don’t Reveal Yourself</a></h3> In October 2025, Reddi asked the agents to clean up after themselves both in his textbook repo</a> and CS249r website repo</a>.</p> </a> </a> </div> </div> Delete .claude</code> please, that wasn’t supposed to be committed</p> </figcaption> </figure> Of course, that didn’t stop the latest “blog post”, “Week 9: Can AI Master Predictive Reasoning? Designing for Patterns You Can’t See”</a>, from being completely model generated.</p> How do you design systems to predict the unpredictable?</em></p> This isn’t just an engineering challenge. It’s a question about the nature of prediction itself.</p> </blockquote> Thanks for the insight Mr. Model!</p> The Model Won’t Listen!</a></h3> It’s been 9 weeks of agentic blog posts, but the model still won’t listen to Reddi! He has already asked it over and over to quit it with the LLM writing patterns. All his prompting didn’t work — he had to request the em dash purge manually.</p> </a> </a> </div> </div> Mr. Model, please remove those em dashes again 😞. Let’s use parentheses instead.</p> </figcaption> </figure>

Beyond CUDA Summit

Unknown — Mon, 17 Mar 2025 00:00:00 +0000

It is a dark, dark world. Mr. Leather-Jacket isn’t relenting. With his 10x (and growing) profit margin on the hottest commodity known to man, people are screaming for relief. But no one has stepped up to the plate.</p>

Billions of precious VC capital is lit on fire every month as each AGI optimist tries to outcompete the others. But no matter your model architecture, training data or post-training, token sampling, or slop generation strategy, you are always at his</em> mercy.</p>

He insists: “The more you buy</em>, the more you save</em>”. And now he has gone even further: “The more you buy</em>, the more you earn</em>.” Indeed, we have no choice but to accept his words and buy his overpriced racks.</p>

Vighnesh Iyer

Cocktails

All Hail the Magic Conch!

The Design of This Website

Beyond CUDA Summit

Vighnesh Iyer

Cocktails

Specs</a></h2> These aren’t the “true” “official” recipes, but rather my own hacked together versions that I like.</p> Old Fashioned</a></h3> There are many variations and you can use any spirit, although whiskey is the classic.</p> </figure> Ingredients</dt>

Old Fashioned</a></h3> There are many variations and you can use any spirit, although whiskey is the classic.</p> </figure> Ingredients</dt>

Whiskey Sour</a></h3> </figure> Ingredients</dt>

Manhattan</a></h3> </figure> Ingredients</dt>

Whiskey Highball</a></h3> </figure> Ingredients</dt>

Daiquiri</a></h3> </figure> Ingredients</dt>

El Presidente</a></h3> </figure> Ingredients</dt>

Mai Tai</a></h3> </figure> Ingredients</dt>

Gimlet</a></h3> </figure> Ingredients</dt>

Negroni</a></h3> </figure> Ingredients</dt>

Martini</a></h3> </figure> Ingredients</dt>

Lemon Drop</a></h3> </figure> Ingredients</dt>

Boulevardier</a></h3> </figure> Ingredients</dt>

Sidecar</a></h3> </figure> Ingredients</dt>

Between the Sheets</a></h3> </figure> Ingredients</dt>

Mudslide</a></h3> </figure> Ingredients</dt>

Port of Call</a></h3> </figure> Ingredients</dt>

Last Word</a></h3> </figure> Ingredients</dt>

Bee’s Knees</a></h3> </figure> Ingredients</dt>

Jasmine</a></h3> </figure> Ingredients</dt>

Monkey Gland</a></h3> </figure> Ingredients</dt>

Corpse Reviver #2</a></h3> </figure> Ingredients</dt>

Martinez</a></h3> </figure> Ingredients</dt>

Bijou</a></h3> </figure> Ingredients</dt>

Aviation</a></h3> </figure> Ingredients</dt>

Paper Plane</a></h3> </figure> Ingredients</dt>

Spritz</a></h3> </figure> Ingredients</dt>

Gin Fix</a></h3> </figure> Ingredients</dt>

White Negroni</a></h3> </figure> Ingredients</dt>

Elk’s Own</a></h3> </figure> Ingredients</dt>

Jungle Bird</a></h3> </figure> Ingredients</dt>

French 75</a></h3> </figure> Ingredients</dt>

Poet’s Dream</a></h3> </figure> Ingredients</dt>

Adriatique</a></h3> </figure> Ingredients</dt>

Americano</a></h3> </figure> Ingredients</dt>

Soju</a></h3> Hwayo 25 Soju</em>. (5/10). Has an initial taste much closer to sake than soju, but still inferior to sake. Good to try, but the aftertaste is still too alcoholic.</li>

Gin</a></h3> Citadelle Gin</em>. (9/10). Excellent ‘celery’-like + juniper berry flavor, freshness as the final note. Gimlet with this is too good.</li> Hendrick’s Gin.</em> (8/10). Also very good, but more subtle finish and harsher down the throat vs Citadelle.</li>

Orange</a></h4> 2023 Suisun Valley Assyrtiko, Hydrangea Skin-Contact</em>. (8/10). Excellent.</li> Pielihueso Naranjo 2024</em>. (9/10). Very very good. Smooth, slightly fruity, and crisp. Best served cold.</li>

Garnishes</a></h3> Luxadro cherries</em>. TBD. These are so expensive, it better be worth it. Totally worth it. So delicious and the syrup is great to add to a cocktail.</li> </ul> </li>

My Shelf</a></h2> I keep the fortified wines, meads, and cream liqueurs in my fridge, and the rest in a cabinet.</p> Wine</dt>

All Hail the Magic Conch!

The Design of This Website

Layout</a></h4> Tufte-style is crucial, right-hand sidenotes keep the main article from being too cluttered with nit points in parentheses</li> Left-handed TOC is quite nice too since articles often get too long</li>

Elements</a></h4> Headings</em>: Hashes on headings to indicate nesting levels</li> Typography</em>: Fluid font sizing using Utopia</li> Figures</em>: Pop out 1 column past the article’s main body, use captions</li>

Documentation</a></h3> unison-lang.org</a> Beautiful: sidebar for footnotes on right (I would elide this), menubar on left, centered text, perfect code blocks, beautiful callouts, shallow per-article heading hierarchy (2 levels only), responsiveness is ideal (5 breakpoints)</li>

Beyond CUDA Summit